1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This transformation implements the well known scalar replacement of
11 /// aggregates transformation. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #include "llvm/Transforms/Scalar.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/AssumptionTracker.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DIBuilder.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstVisitor.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Operator.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 #define DEBUG_TYPE "sroa"
69 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
70 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
71 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
72 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
73 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
74 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
75 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
76 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
77 STATISTIC(NumDeleted, "Number of instructions deleted");
78 STATISTIC(NumVectorized, "Number of vectorized aggregates");
80 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
81 /// forming SSA values through the SSAUpdater infrastructure.
82 static cl::opt<bool> ForceSSAUpdater("force-ssa-updater", cl::init(false),
85 /// Hidden option to enable randomly shuffling the slices to help uncover
86 /// instability in their order.
87 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
88 cl::init(false), cl::Hidden);
90 /// Hidden option to experiment with completely strict handling of inbounds
92 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
96 /// \brief A custom IRBuilder inserter which prefixes all names if they are
98 template <bool preserveNames = true>
99 class IRBuilderPrefixedInserter
100 : public IRBuilderDefaultInserter<preserveNames> {
104 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
107 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
108 BasicBlock::iterator InsertPt) const {
109 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
110 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
114 // Specialization for not preserving the name is trivial.
116 class IRBuilderPrefixedInserter<false>
117 : public IRBuilderDefaultInserter<false> {
119 void SetNamePrefix(const Twine &P) {}
122 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
124 typedef llvm::IRBuilder<true, ConstantFolder, IRBuilderPrefixedInserter<true>>
127 typedef llvm::IRBuilder<false, ConstantFolder, IRBuilderPrefixedInserter<false>>
133 /// \brief A used slice of an alloca.
135 /// This structure represents a slice of an alloca used by some instruction. It
136 /// stores both the begin and end offsets of this use, a pointer to the use
137 /// itself, and a flag indicating whether we can classify the use as splittable
138 /// or not when forming partitions of the alloca.
140 /// \brief The beginning offset of the range.
141 uint64_t BeginOffset;
143 /// \brief The ending offset, not included in the range.
146 /// \brief Storage for both the use of this slice and whether it can be
148 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
151 Slice() : BeginOffset(), EndOffset() {}
152 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
153 : BeginOffset(BeginOffset), EndOffset(EndOffset),
154 UseAndIsSplittable(U, IsSplittable) {}
156 uint64_t beginOffset() const { return BeginOffset; }
157 uint64_t endOffset() const { return EndOffset; }
159 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
160 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
162 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
164 bool isDead() const { return getUse() == nullptr; }
165 void kill() { UseAndIsSplittable.setPointer(nullptr); }
167 /// \brief Support for ordering ranges.
169 /// This provides an ordering over ranges such that start offsets are
170 /// always increasing, and within equal start offsets, the end offsets are
171 /// decreasing. Thus the spanning range comes first in a cluster with the
172 /// same start position.
173 bool operator<(const Slice &RHS) const {
174 if (beginOffset() < RHS.beginOffset())
176 if (beginOffset() > RHS.beginOffset())
178 if (isSplittable() != RHS.isSplittable())
179 return !isSplittable();
180 if (endOffset() > RHS.endOffset())
185 /// \brief Support comparison with a single offset to allow binary searches.
186 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
187 uint64_t RHSOffset) {
188 return LHS.beginOffset() < RHSOffset;
190 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
192 return LHSOffset < RHS.beginOffset();
195 bool operator==(const Slice &RHS) const {
196 return isSplittable() == RHS.isSplittable() &&
197 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
199 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
201 } // end anonymous namespace
204 template <typename T> struct isPodLike;
205 template <> struct isPodLike<Slice> { static const bool value = true; };
209 /// \brief Representation of the alloca slices.
211 /// This class represents the slices of an alloca which are formed by its
212 /// various uses. If a pointer escapes, we can't fully build a representation
213 /// for the slices used and we reflect that in this structure. The uses are
214 /// stored, sorted by increasing beginning offset and with unsplittable slices
215 /// starting at a particular offset before splittable slices.
218 /// \brief Construct the slices of a particular alloca.
219 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
221 /// \brief Test whether a pointer to the allocation escapes our analysis.
223 /// If this is true, the slices are never fully built and should be
225 bool isEscaped() const { return PointerEscapingInstr; }
227 /// \brief Support for iterating over the slices.
229 typedef SmallVectorImpl<Slice>::iterator iterator;
230 typedef iterator_range<iterator> range;
231 iterator begin() { return Slices.begin(); }
232 iterator end() { return Slices.end(); }
234 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
235 typedef iterator_range<const_iterator> const_range;
236 const_iterator begin() const { return Slices.begin(); }
237 const_iterator end() const { return Slices.end(); }
240 /// \brief Access the dead users for this alloca.
241 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
243 /// \brief Access the dead operands referring to this alloca.
245 /// These are operands which have cannot actually be used to refer to the
246 /// alloca as they are outside its range and the user doesn't correct for
247 /// that. These mostly consist of PHI node inputs and the like which we just
248 /// need to replace with undef.
249 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
251 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
252 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
253 void printSlice(raw_ostream &OS, const_iterator I,
254 StringRef Indent = " ") const;
255 void printUse(raw_ostream &OS, const_iterator I,
256 StringRef Indent = " ") const;
257 void print(raw_ostream &OS) const;
258 void dump(const_iterator I) const;
263 template <typename DerivedT, typename RetT = void> class BuilderBase;
265 friend class AllocaSlices::SliceBuilder;
267 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
268 /// \brief Handle to alloca instruction to simplify method interfaces.
272 /// \brief The instruction responsible for this alloca not having a known set
275 /// When an instruction (potentially) escapes the pointer to the alloca, we
276 /// store a pointer to that here and abort trying to form slices of the
277 /// alloca. This will be null if the alloca slices are analyzed successfully.
278 Instruction *PointerEscapingInstr;
280 /// \brief The slices of the alloca.
282 /// We store a vector of the slices formed by uses of the alloca here. This
283 /// vector is sorted by increasing begin offset, and then the unsplittable
284 /// slices before the splittable ones. See the Slice inner class for more
286 SmallVector<Slice, 8> Slices;
288 /// \brief Instructions which will become dead if we rewrite the alloca.
290 /// Note that these are not separated by slice. This is because we expect an
291 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
292 /// all these instructions can simply be removed and replaced with undef as
293 /// they come from outside of the allocated space.
294 SmallVector<Instruction *, 8> DeadUsers;
296 /// \brief Operands which will become dead if we rewrite the alloca.
298 /// These are operands that in their particular use can be replaced with
299 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
300 /// to PHI nodes and the like. They aren't entirely dead (there might be
301 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
302 /// want to swap this particular input for undef to simplify the use lists of
304 SmallVector<Use *, 8> DeadOperands;
308 static Value *foldSelectInst(SelectInst &SI) {
309 // If the condition being selected on is a constant or the same value is
310 // being selected between, fold the select. Yes this does (rarely) happen
312 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
313 return SI.getOperand(1 + CI->isZero());
314 if (SI.getOperand(1) == SI.getOperand(2))
315 return SI.getOperand(1);
320 /// \brief A helper that folds a PHI node or a select.
321 static Value *foldPHINodeOrSelectInst(Instruction &I) {
322 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
323 // If PN merges together the same value, return that value.
324 return PN->hasConstantValue();
326 return foldSelectInst(cast<SelectInst>(I));
329 /// \brief Builder for the alloca slices.
331 /// This class builds a set of alloca slices by recursively visiting the uses
332 /// of an alloca and making a slice for each load and store at each offset.
333 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
334 friend class PtrUseVisitor<SliceBuilder>;
335 friend class InstVisitor<SliceBuilder>;
336 typedef PtrUseVisitor<SliceBuilder> Base;
338 const uint64_t AllocSize;
341 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
342 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
344 /// \brief Set to de-duplicate dead instructions found in the use walk.
345 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
348 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
349 : PtrUseVisitor<SliceBuilder>(DL),
350 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
353 void markAsDead(Instruction &I) {
354 if (VisitedDeadInsts.insert(&I).second)
355 AS.DeadUsers.push_back(&I);
358 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
359 bool IsSplittable = false) {
360 // Completely skip uses which have a zero size or start either before or
361 // past the end of the allocation.
362 if (Size == 0 || Offset.uge(AllocSize)) {
363 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
364 << " which has zero size or starts outside of the "
365 << AllocSize << " byte alloca:\n"
366 << " alloca: " << AS.AI << "\n"
367 << " use: " << I << "\n");
368 return markAsDead(I);
371 uint64_t BeginOffset = Offset.getZExtValue();
372 uint64_t EndOffset = BeginOffset + Size;
374 // Clamp the end offset to the end of the allocation. Note that this is
375 // formulated to handle even the case where "BeginOffset + Size" overflows.
376 // This may appear superficially to be something we could ignore entirely,
377 // but that is not so! There may be widened loads or PHI-node uses where
378 // some instructions are dead but not others. We can't completely ignore
379 // them, and so have to record at least the information here.
380 assert(AllocSize >= BeginOffset); // Established above.
381 if (Size > AllocSize - BeginOffset) {
382 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
383 << " to remain within the " << AllocSize << " byte alloca:\n"
384 << " alloca: " << AS.AI << "\n"
385 << " use: " << I << "\n");
386 EndOffset = AllocSize;
389 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
392 void visitBitCastInst(BitCastInst &BC) {
394 return markAsDead(BC);
396 return Base::visitBitCastInst(BC);
399 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
400 if (GEPI.use_empty())
401 return markAsDead(GEPI);
403 if (SROAStrictInbounds && GEPI.isInBounds()) {
404 // FIXME: This is a manually un-factored variant of the basic code inside
405 // of GEPs with checking of the inbounds invariant specified in the
406 // langref in a very strict sense. If we ever want to enable
407 // SROAStrictInbounds, this code should be factored cleanly into
408 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
409 // by writing out the code here where we have tho underlying allocation
410 // size readily available.
411 APInt GEPOffset = Offset;
412 for (gep_type_iterator GTI = gep_type_begin(GEPI),
413 GTE = gep_type_end(GEPI);
415 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
419 // Handle a struct index, which adds its field offset to the pointer.
420 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
421 unsigned ElementIdx = OpC->getZExtValue();
422 const StructLayout *SL = DL.getStructLayout(STy);
424 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
426 // For array or vector indices, scale the index by the size of the
428 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
429 GEPOffset += Index * APInt(Offset.getBitWidth(),
430 DL.getTypeAllocSize(GTI.getIndexedType()));
433 // If this index has computed an intermediate pointer which is not
434 // inbounds, then the result of the GEP is a poison value and we can
435 // delete it and all uses.
436 if (GEPOffset.ugt(AllocSize))
437 return markAsDead(GEPI);
441 return Base::visitGetElementPtrInst(GEPI);
444 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
445 uint64_t Size, bool IsVolatile) {
446 // We allow splitting of loads and stores where the type is an integer type
447 // and cover the entire alloca. This prevents us from splitting over
449 // FIXME: In the great blue eventually, we should eagerly split all integer
450 // loads and stores, and then have a separate step that merges adjacent
451 // alloca partitions into a single partition suitable for integer widening.
452 // Or we should skip the merge step and rely on GVN and other passes to
453 // merge adjacent loads and stores that survive mem2reg.
455 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
457 insertUse(I, Offset, Size, IsSplittable);
460 void visitLoadInst(LoadInst &LI) {
461 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
462 "All simple FCA loads should have been pre-split");
465 return PI.setAborted(&LI);
467 uint64_t Size = DL.getTypeStoreSize(LI.getType());
468 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
471 void visitStoreInst(StoreInst &SI) {
472 Value *ValOp = SI.getValueOperand();
474 return PI.setEscapedAndAborted(&SI);
476 return PI.setAborted(&SI);
478 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
480 // If this memory access can be shown to *statically* extend outside the
481 // bounds of of the allocation, it's behavior is undefined, so simply
482 // ignore it. Note that this is more strict than the generic clamping
483 // behavior of insertUse. We also try to handle cases which might run the
485 // FIXME: We should instead consider the pointer to have escaped if this
486 // function is being instrumented for addressing bugs or race conditions.
487 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
488 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
489 << " which extends past the end of the " << AllocSize
491 << " alloca: " << AS.AI << "\n"
492 << " use: " << SI << "\n");
493 return markAsDead(SI);
496 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
497 "All simple FCA stores should have been pre-split");
498 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
501 void visitMemSetInst(MemSetInst &II) {
502 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
503 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
504 if ((Length && Length->getValue() == 0) ||
505 (IsOffsetKnown && Offset.uge(AllocSize)))
506 // Zero-length mem transfer intrinsics can be ignored entirely.
507 return markAsDead(II);
510 return PI.setAborted(&II);
512 insertUse(II, Offset, Length ? Length->getLimitedValue()
513 : AllocSize - Offset.getLimitedValue(),
517 void visitMemTransferInst(MemTransferInst &II) {
518 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
519 if (Length && Length->getValue() == 0)
520 // Zero-length mem transfer intrinsics can be ignored entirely.
521 return markAsDead(II);
523 // Because we can visit these intrinsics twice, also check to see if the
524 // first time marked this instruction as dead. If so, skip it.
525 if (VisitedDeadInsts.count(&II))
529 return PI.setAborted(&II);
531 // This side of the transfer is completely out-of-bounds, and so we can
532 // nuke the entire transfer. However, we also need to nuke the other side
533 // if already added to our partitions.
534 // FIXME: Yet another place we really should bypass this when
535 // instrumenting for ASan.
536 if (Offset.uge(AllocSize)) {
537 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
538 MemTransferSliceMap.find(&II);
539 if (MTPI != MemTransferSliceMap.end())
540 AS.Slices[MTPI->second].kill();
541 return markAsDead(II);
544 uint64_t RawOffset = Offset.getLimitedValue();
545 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
547 // Check for the special case where the same exact value is used for both
549 if (*U == II.getRawDest() && *U == II.getRawSource()) {
550 // For non-volatile transfers this is a no-op.
551 if (!II.isVolatile())
552 return markAsDead(II);
554 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
557 // If we have seen both source and destination for a mem transfer, then
558 // they both point to the same alloca.
560 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
561 std::tie(MTPI, Inserted) =
562 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
563 unsigned PrevIdx = MTPI->second;
565 Slice &PrevP = AS.Slices[PrevIdx];
567 // Check if the begin offsets match and this is a non-volatile transfer.
568 // In that case, we can completely elide the transfer.
569 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
571 return markAsDead(II);
574 // Otherwise we have an offset transfer within the same alloca. We can't
576 PrevP.makeUnsplittable();
579 // Insert the use now that we've fixed up the splittable nature.
580 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
582 // Check that we ended up with a valid index in the map.
583 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
584 "Map index doesn't point back to a slice with this user.");
587 // Disable SRoA for any intrinsics except for lifetime invariants.
588 // FIXME: What about debug intrinsics? This matches old behavior, but
589 // doesn't make sense.
590 void visitIntrinsicInst(IntrinsicInst &II) {
592 return PI.setAborted(&II);
594 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
595 II.getIntrinsicID() == Intrinsic::lifetime_end) {
596 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
597 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
598 Length->getLimitedValue());
599 insertUse(II, Offset, Size, true);
603 Base::visitIntrinsicInst(II);
606 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
607 // We consider any PHI or select that results in a direct load or store of
608 // the same offset to be a viable use for slicing purposes. These uses
609 // are considered unsplittable and the size is the maximum loaded or stored
611 SmallPtrSet<Instruction *, 4> Visited;
612 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
613 Visited.insert(Root);
614 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
615 // If there are no loads or stores, the access is dead. We mark that as
616 // a size zero access.
619 Instruction *I, *UsedI;
620 std::tie(UsedI, I) = Uses.pop_back_val();
622 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
623 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
626 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
627 Value *Op = SI->getOperand(0);
630 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
634 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
635 if (!GEP->hasAllZeroIndices())
637 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
638 !isa<SelectInst>(I)) {
642 for (User *U : I->users())
643 if (Visited.insert(cast<Instruction>(U)).second)
644 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
645 } while (!Uses.empty());
650 void visitPHINodeOrSelectInst(Instruction &I) {
651 assert(isa<PHINode>(I) || isa<SelectInst>(I));
653 return markAsDead(I);
655 // TODO: We could use SimplifyInstruction here to fold PHINodes and
656 // SelectInsts. However, doing so requires to change the current
657 // dead-operand-tracking mechanism. For instance, suppose neither loading
658 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
659 // trap either. However, if we simply replace %U with undef using the
660 // current dead-operand-tracking mechanism, "load (select undef, undef,
661 // %other)" may trap because the select may return the first operand
663 if (Value *Result = foldPHINodeOrSelectInst(I)) {
665 // If the result of the constant fold will be the pointer, recurse
666 // through the PHI/select as if we had RAUW'ed it.
669 // Otherwise the operand to the PHI/select is dead, and we can replace
671 AS.DeadOperands.push_back(U);
677 return PI.setAborted(&I);
679 // See if we already have computed info on this node.
680 uint64_t &Size = PHIOrSelectSizes[&I];
682 // This is a new PHI/Select, check for an unsafe use of it.
683 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
684 return PI.setAborted(UnsafeI);
687 // For PHI and select operands outside the alloca, we can't nuke the entire
688 // phi or select -- the other side might still be relevant, so we special
689 // case them here and use a separate structure to track the operands
690 // themselves which should be replaced with undef.
691 // FIXME: This should instead be escaped in the event we're instrumenting
692 // for address sanitization.
693 if (Offset.uge(AllocSize)) {
694 AS.DeadOperands.push_back(U);
698 insertUse(I, Offset, Size);
701 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
703 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
705 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
706 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
709 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
711 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
714 PointerEscapingInstr(nullptr) {
715 SliceBuilder PB(DL, AI, *this);
716 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
717 if (PtrI.isEscaped() || PtrI.isAborted()) {
718 // FIXME: We should sink the escape vs. abort info into the caller nicely,
719 // possibly by just storing the PtrInfo in the AllocaSlices.
720 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
721 : PtrI.getAbortingInst();
722 assert(PointerEscapingInstr && "Did not track a bad instruction");
726 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
732 #if __cplusplus >= 201103L && !defined(NDEBUG)
733 if (SROARandomShuffleSlices) {
734 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
735 std::shuffle(Slices.begin(), Slices.end(), MT);
739 // Sort the uses. This arranges for the offsets to be in ascending order,
740 // and the sizes to be in descending order.
741 std::sort(Slices.begin(), Slices.end());
744 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
746 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
747 StringRef Indent) const {
748 printSlice(OS, I, Indent);
749 printUse(OS, I, Indent);
752 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
753 StringRef Indent) const {
754 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
755 << " slice #" << (I - begin())
756 << (I->isSplittable() ? " (splittable)" : "") << "\n";
759 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
760 StringRef Indent) const {
761 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
764 void AllocaSlices::print(raw_ostream &OS) const {
765 if (PointerEscapingInstr) {
766 OS << "Can't analyze slices for alloca: " << AI << "\n"
767 << " A pointer to this alloca escaped by:\n"
768 << " " << *PointerEscapingInstr << "\n";
772 OS << "Slices of alloca: " << AI << "\n";
773 for (const_iterator I = begin(), E = end(); I != E; ++I)
777 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
780 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
782 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
785 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
787 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
788 /// the loads and stores of an alloca instruction, as well as updating its
789 /// debug information. This is used when a domtree is unavailable and thus
790 /// mem2reg in its full form can't be used to handle promotion of allocas to
792 class AllocaPromoter : public LoadAndStorePromoter {
796 SmallVector<DbgDeclareInst *, 4> DDIs;
797 SmallVector<DbgValueInst *, 4> DVIs;
800 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
801 AllocaInst &AI, DIBuilder &DIB)
802 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
804 void run(const SmallVectorImpl<Instruction *> &Insts) {
805 // Retain the debug information attached to the alloca for use when
806 // rewriting loads and stores.
807 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
808 if (auto *DebugNode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
809 for (User *U : DebugNode->users())
810 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
812 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
817 LoadAndStorePromoter::run(Insts);
819 // While we have the debug information, clear it off of the alloca. The
820 // caller takes care of deleting the alloca.
821 while (!DDIs.empty())
822 DDIs.pop_back_val()->eraseFromParent();
823 while (!DVIs.empty())
824 DVIs.pop_back_val()->eraseFromParent();
828 isInstInList(Instruction *I,
829 const SmallVectorImpl<Instruction *> &Insts) const override {
831 if (LoadInst *LI = dyn_cast<LoadInst>(I))
832 Ptr = LI->getOperand(0);
834 Ptr = cast<StoreInst>(I)->getPointerOperand();
836 // Only used to detect cycles, which will be rare and quickly found as
837 // we're walking up a chain of defs rather than down through uses.
838 SmallPtrSet<Value *, 4> Visited;
844 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
845 Ptr = BCI->getOperand(0);
846 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
847 Ptr = GEPI->getPointerOperand();
851 } while (Visited.insert(Ptr).second);
856 void updateDebugInfo(Instruction *Inst) const override {
857 for (DbgDeclareInst *DDI : DDIs)
858 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
859 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
860 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
861 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
862 for (DbgValueInst *DVI : DVIs) {
863 Value *Arg = nullptr;
864 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
865 // If an argument is zero extended then use argument directly. The ZExt
866 // may be zapped by an optimization pass in future.
867 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
868 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
869 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
870 Arg = dyn_cast<Argument>(SExt->getOperand(0));
872 Arg = SI->getValueOperand();
873 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
874 Arg = LI->getPointerOperand();
878 Instruction *DbgVal =
879 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
880 DIExpression(DVI->getExpression()), Inst);
881 DbgVal->setDebugLoc(DVI->getDebugLoc());
885 } // end anon namespace
888 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
890 /// This pass takes allocations which can be completely analyzed (that is, they
891 /// don't escape) and tries to turn them into scalar SSA values. There are
892 /// a few steps to this process.
894 /// 1) It takes allocations of aggregates and analyzes the ways in which they
895 /// are used to try to split them into smaller allocations, ideally of
896 /// a single scalar data type. It will split up memcpy and memset accesses
897 /// as necessary and try to isolate individual scalar accesses.
898 /// 2) It will transform accesses into forms which are suitable for SSA value
899 /// promotion. This can be replacing a memset with a scalar store of an
900 /// integer value, or it can involve speculating operations on a PHI or
901 /// select to be a PHI or select of the results.
902 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
903 /// onto insert and extract operations on a vector value, and convert them to
904 /// this form. By doing so, it will enable promotion of vector aggregates to
905 /// SSA vector values.
906 class SROA : public FunctionPass {
907 const bool RequiresDomTree;
910 const DataLayout *DL;
912 AssumptionTracker *AT;
914 /// \brief Worklist of alloca instructions to simplify.
916 /// Each alloca in the function is added to this. Each new alloca formed gets
917 /// added to it as well to recursively simplify unless that alloca can be
918 /// directly promoted. Finally, each time we rewrite a use of an alloca other
919 /// the one being actively rewritten, we add it back onto the list if not
920 /// already present to ensure it is re-visited.
921 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
923 /// \brief A collection of instructions to delete.
924 /// We try to batch deletions to simplify code and make things a bit more
926 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
928 /// \brief Post-promotion worklist.
930 /// Sometimes we discover an alloca which has a high probability of becoming
931 /// viable for SROA after a round of promotion takes place. In those cases,
932 /// the alloca is enqueued here for re-processing.
934 /// Note that we have to be very careful to clear allocas out of this list in
935 /// the event they are deleted.
936 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
938 /// \brief A collection of alloca instructions we can directly promote.
939 std::vector<AllocaInst *> PromotableAllocas;
941 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
943 /// All of these PHIs have been checked for the safety of speculation and by
944 /// being speculated will allow promoting allocas currently in the promotable
946 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
948 /// \brief A worklist of select instructions to speculate prior to promoting
951 /// All of these select instructions have been checked for the safety of
952 /// speculation and by being speculated will allow promoting allocas
953 /// currently in the promotable queue.
954 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
956 /// Debug intrinsics do not show up as regular uses in the
957 /// IR. This side-table holds the missing use edges.
958 DenseMap<AllocaInst *, DbgDeclareInst *> DbgDeclares;
961 SROA(bool RequiresDomTree = true)
962 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
963 DL(nullptr), DT(nullptr) {
964 initializeSROAPass(*PassRegistry::getPassRegistry());
966 bool runOnFunction(Function &F) override;
967 void getAnalysisUsage(AnalysisUsage &AU) const override;
969 const char *getPassName() const override { return "SROA"; }
973 friend class PHIOrSelectSpeculator;
974 friend class AllocaSliceRewriter;
976 bool rewritePartition(AllocaInst &AI, AllocaSlices &AS,
977 AllocaSlices::iterator B, AllocaSlices::iterator E,
978 int64_t BeginOffset, int64_t EndOffset,
979 ArrayRef<AllocaSlices::iterator> SplitUses);
980 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
981 bool runOnAlloca(AllocaInst &AI);
982 void clobberUse(Use &U);
983 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
984 bool promoteAllocas(Function &F);
990 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
991 return new SROA(RequiresDomTree);
994 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
996 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
997 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
998 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1001 /// Walk the range of a partitioning looking for a common type to cover this
1002 /// sequence of slices.
1003 static Type *findCommonType(AllocaSlices::const_iterator B,
1004 AllocaSlices::const_iterator E,
1005 uint64_t EndOffset) {
1007 bool TyIsCommon = true;
1008 IntegerType *ITy = nullptr;
1010 // Note that we need to look at *every* alloca slice's Use to ensure we
1011 // always get consistent results regardless of the order of slices.
1012 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1013 Use *U = I->getUse();
1014 if (isa<IntrinsicInst>(*U->getUser()))
1016 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1019 Type *UserTy = nullptr;
1020 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1021 UserTy = LI->getType();
1022 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1023 UserTy = SI->getValueOperand()->getType();
1026 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1027 // If the type is larger than the partition, skip it. We only encounter
1028 // this for split integer operations where we want to use the type of the
1029 // entity causing the split. Also skip if the type is not a byte width
1031 if (UserITy->getBitWidth() % 8 != 0 ||
1032 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1035 // Track the largest bitwidth integer type used in this way in case there
1036 // is no common type.
1037 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1041 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1042 // depend on types skipped above.
1043 if (!UserTy || (Ty && Ty != UserTy))
1044 TyIsCommon = false; // Give up on anything but an iN type.
1049 return TyIsCommon ? Ty : ITy;
1052 /// PHI instructions that use an alloca and are subsequently loaded can be
1053 /// rewritten to load both input pointers in the pred blocks and then PHI the
1054 /// results, allowing the load of the alloca to be promoted.
1056 /// %P2 = phi [i32* %Alloca, i32* %Other]
1057 /// %V = load i32* %P2
1059 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1061 /// %V2 = load i32* %Other
1063 /// %V = phi [i32 %V1, i32 %V2]
1065 /// We can do this to a select if its only uses are loads and if the operands
1066 /// to the select can be loaded unconditionally.
1068 /// FIXME: This should be hoisted into a generic utility, likely in
1069 /// Transforms/Util/Local.h
1070 static bool isSafePHIToSpeculate(PHINode &PN, const DataLayout *DL = nullptr) {
1071 // For now, we can only do this promotion if the load is in the same block
1072 // as the PHI, and if there are no stores between the phi and load.
1073 // TODO: Allow recursive phi users.
1074 // TODO: Allow stores.
1075 BasicBlock *BB = PN.getParent();
1076 unsigned MaxAlign = 0;
1077 bool HaveLoad = false;
1078 for (User *U : PN.users()) {
1079 LoadInst *LI = dyn_cast<LoadInst>(U);
1080 if (!LI || !LI->isSimple())
1083 // For now we only allow loads in the same block as the PHI. This is
1084 // a common case that happens when instcombine merges two loads through
1086 if (LI->getParent() != BB)
1089 // Ensure that there are no instructions between the PHI and the load that
1091 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1092 if (BBI->mayWriteToMemory())
1095 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1102 // We can only transform this if it is safe to push the loads into the
1103 // predecessor blocks. The only thing to watch out for is that we can't put
1104 // a possibly trapping load in the predecessor if it is a critical edge.
1105 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1106 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1107 Value *InVal = PN.getIncomingValue(Idx);
1109 // If the value is produced by the terminator of the predecessor (an
1110 // invoke) or it has side-effects, there is no valid place to put a load
1111 // in the predecessor.
1112 if (TI == InVal || TI->mayHaveSideEffects())
1115 // If the predecessor has a single successor, then the edge isn't
1117 if (TI->getNumSuccessors() == 1)
1120 // If this pointer is always safe to load, or if we can prove that there
1121 // is already a load in the block, then we can move the load to the pred
1123 if (InVal->isDereferenceablePointer(DL) ||
1124 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1133 static void speculatePHINodeLoads(PHINode &PN) {
1134 DEBUG(dbgs() << " original: " << PN << "\n");
1136 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1137 IRBuilderTy PHIBuilder(&PN);
1138 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1139 PN.getName() + ".sroa.speculated");
1141 // Get the AA tags and alignment to use from one of the loads. It doesn't
1142 // matter which one we get and if any differ.
1143 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1146 SomeLoad->getAAMetadata(AATags);
1147 unsigned Align = SomeLoad->getAlignment();
1149 // Rewrite all loads of the PN to use the new PHI.
1150 while (!PN.use_empty()) {
1151 LoadInst *LI = cast<LoadInst>(PN.user_back());
1152 LI->replaceAllUsesWith(NewPN);
1153 LI->eraseFromParent();
1156 // Inject loads into all of the pred blocks.
1157 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1158 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1159 TerminatorInst *TI = Pred->getTerminator();
1160 Value *InVal = PN.getIncomingValue(Idx);
1161 IRBuilderTy PredBuilder(TI);
1163 LoadInst *Load = PredBuilder.CreateLoad(
1164 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1165 ++NumLoadsSpeculated;
1166 Load->setAlignment(Align);
1168 Load->setAAMetadata(AATags);
1169 NewPN->addIncoming(Load, Pred);
1172 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1173 PN.eraseFromParent();
1176 /// Select instructions that use an alloca and are subsequently loaded can be
1177 /// rewritten to load both input pointers and then select between the result,
1178 /// allowing the load of the alloca to be promoted.
1180 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1181 /// %V = load i32* %P2
1183 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1184 /// %V2 = load i32* %Other
1185 /// %V = select i1 %cond, i32 %V1, i32 %V2
1187 /// We can do this to a select if its only uses are loads and if the operand
1188 /// to the select can be loaded unconditionally.
1189 static bool isSafeSelectToSpeculate(SelectInst &SI,
1190 const DataLayout *DL = nullptr) {
1191 Value *TValue = SI.getTrueValue();
1192 Value *FValue = SI.getFalseValue();
1193 bool TDerefable = TValue->isDereferenceablePointer(DL);
1194 bool FDerefable = FValue->isDereferenceablePointer(DL);
1196 for (User *U : SI.users()) {
1197 LoadInst *LI = dyn_cast<LoadInst>(U);
1198 if (!LI || !LI->isSimple())
1201 // Both operands to the select need to be dereferencable, either
1202 // absolutely (e.g. allocas) or at this point because we can see other
1205 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1208 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1215 static void speculateSelectInstLoads(SelectInst &SI) {
1216 DEBUG(dbgs() << " original: " << SI << "\n");
1218 IRBuilderTy IRB(&SI);
1219 Value *TV = SI.getTrueValue();
1220 Value *FV = SI.getFalseValue();
1221 // Replace the loads of the select with a select of two loads.
1222 while (!SI.use_empty()) {
1223 LoadInst *LI = cast<LoadInst>(SI.user_back());
1224 assert(LI->isSimple() && "We only speculate simple loads");
1226 IRB.SetInsertPoint(LI);
1228 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1230 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1231 NumLoadsSpeculated += 2;
1233 // Transfer alignment and AA info if present.
1234 TL->setAlignment(LI->getAlignment());
1235 FL->setAlignment(LI->getAlignment());
1238 LI->getAAMetadata(Tags);
1240 TL->setAAMetadata(Tags);
1241 FL->setAAMetadata(Tags);
1244 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1245 LI->getName() + ".sroa.speculated");
1247 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1248 LI->replaceAllUsesWith(V);
1249 LI->eraseFromParent();
1251 SI.eraseFromParent();
1254 /// \brief Build a GEP out of a base pointer and indices.
1256 /// This will return the BasePtr if that is valid, or build a new GEP
1257 /// instruction using the IRBuilder if GEP-ing is needed.
1258 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1259 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1260 if (Indices.empty())
1263 // A single zero index is a no-op, so check for this and avoid building a GEP
1265 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1268 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1271 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1272 /// TargetTy without changing the offset of the pointer.
1274 /// This routine assumes we've already established a properly offset GEP with
1275 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1276 /// zero-indices down through type layers until we find one the same as
1277 /// TargetTy. If we can't find one with the same type, we at least try to use
1278 /// one with the same size. If none of that works, we just produce the GEP as
1279 /// indicated by Indices to have the correct offset.
1280 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1281 Value *BasePtr, Type *Ty, Type *TargetTy,
1282 SmallVectorImpl<Value *> &Indices,
1285 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1287 // Pointer size to use for the indices.
1288 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1290 // See if we can descend into a struct and locate a field with the correct
1292 unsigned NumLayers = 0;
1293 Type *ElementTy = Ty;
1295 if (ElementTy->isPointerTy())
1298 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1299 ElementTy = ArrayTy->getElementType();
1300 Indices.push_back(IRB.getIntN(PtrSize, 0));
1301 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1302 ElementTy = VectorTy->getElementType();
1303 Indices.push_back(IRB.getInt32(0));
1304 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1305 if (STy->element_begin() == STy->element_end())
1306 break; // Nothing left to descend into.
1307 ElementTy = *STy->element_begin();
1308 Indices.push_back(IRB.getInt32(0));
1313 } while (ElementTy != TargetTy);
1314 if (ElementTy != TargetTy)
1315 Indices.erase(Indices.end() - NumLayers, Indices.end());
1317 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1320 /// \brief Recursively compute indices for a natural GEP.
1322 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1323 /// element types adding appropriate indices for the GEP.
1324 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1325 Value *Ptr, Type *Ty, APInt &Offset,
1327 SmallVectorImpl<Value *> &Indices,
1330 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1333 // We can't recurse through pointer types.
1334 if (Ty->isPointerTy())
1337 // We try to analyze GEPs over vectors here, but note that these GEPs are
1338 // extremely poorly defined currently. The long-term goal is to remove GEPing
1339 // over a vector from the IR completely.
1340 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1341 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1342 if (ElementSizeInBits % 8 != 0) {
1343 // GEPs over non-multiple of 8 size vector elements are invalid.
1346 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1347 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1348 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1350 Offset -= NumSkippedElements * ElementSize;
1351 Indices.push_back(IRB.getInt(NumSkippedElements));
1352 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1353 Offset, TargetTy, Indices, NamePrefix);
1356 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1357 Type *ElementTy = ArrTy->getElementType();
1358 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1359 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1360 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1363 Offset -= NumSkippedElements * ElementSize;
1364 Indices.push_back(IRB.getInt(NumSkippedElements));
1365 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1366 Indices, NamePrefix);
1369 StructType *STy = dyn_cast<StructType>(Ty);
1373 const StructLayout *SL = DL.getStructLayout(STy);
1374 uint64_t StructOffset = Offset.getZExtValue();
1375 if (StructOffset >= SL->getSizeInBytes())
1377 unsigned Index = SL->getElementContainingOffset(StructOffset);
1378 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1379 Type *ElementTy = STy->getElementType(Index);
1380 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1381 return nullptr; // The offset points into alignment padding.
1383 Indices.push_back(IRB.getInt32(Index));
1384 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1385 Indices, NamePrefix);
1388 /// \brief Get a natural GEP from a base pointer to a particular offset and
1389 /// resulting in a particular type.
1391 /// The goal is to produce a "natural" looking GEP that works with the existing
1392 /// composite types to arrive at the appropriate offset and element type for
1393 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1394 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1395 /// Indices, and setting Ty to the result subtype.
1397 /// If no natural GEP can be constructed, this function returns null.
1398 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1399 Value *Ptr, APInt Offset, Type *TargetTy,
1400 SmallVectorImpl<Value *> &Indices,
1402 PointerType *Ty = cast<PointerType>(Ptr->getType());
1404 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1406 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1409 Type *ElementTy = Ty->getElementType();
1410 if (!ElementTy->isSized())
1411 return nullptr; // We can't GEP through an unsized element.
1412 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1413 if (ElementSize == 0)
1414 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1415 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1417 Offset -= NumSkippedElements * ElementSize;
1418 Indices.push_back(IRB.getInt(NumSkippedElements));
1419 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1420 Indices, NamePrefix);
1423 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1424 /// resulting pointer has PointerTy.
1426 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1427 /// and produces the pointer type desired. Where it cannot, it will try to use
1428 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1429 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1430 /// bitcast to the type.
1432 /// The strategy for finding the more natural GEPs is to peel off layers of the
1433 /// pointer, walking back through bit casts and GEPs, searching for a base
1434 /// pointer from which we can compute a natural GEP with the desired
1435 /// properties. The algorithm tries to fold as many constant indices into
1436 /// a single GEP as possible, thus making each GEP more independent of the
1437 /// surrounding code.
1438 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1439 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1440 // Even though we don't look through PHI nodes, we could be called on an
1441 // instruction in an unreachable block, which may be on a cycle.
1442 SmallPtrSet<Value *, 4> Visited;
1443 Visited.insert(Ptr);
1444 SmallVector<Value *, 4> Indices;
1446 // We may end up computing an offset pointer that has the wrong type. If we
1447 // never are able to compute one directly that has the correct type, we'll
1448 // fall back to it, so keep it around here.
1449 Value *OffsetPtr = nullptr;
1451 // Remember any i8 pointer we come across to re-use if we need to do a raw
1453 Value *Int8Ptr = nullptr;
1454 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1456 Type *TargetTy = PointerTy->getPointerElementType();
1459 // First fold any existing GEPs into the offset.
1460 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1461 APInt GEPOffset(Offset.getBitWidth(), 0);
1462 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1464 Offset += GEPOffset;
1465 Ptr = GEP->getPointerOperand();
1466 if (!Visited.insert(Ptr).second)
1470 // See if we can perform a natural GEP here.
1472 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1473 Indices, NamePrefix)) {
1474 if (P->getType() == PointerTy) {
1475 // Zap any offset pointer that we ended up computing in previous rounds.
1476 if (OffsetPtr && OffsetPtr->use_empty())
1477 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1478 I->eraseFromParent();
1486 // Stash this pointer if we've found an i8*.
1487 if (Ptr->getType()->isIntegerTy(8)) {
1489 Int8PtrOffset = Offset;
1492 // Peel off a layer of the pointer and update the offset appropriately.
1493 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1494 Ptr = cast<Operator>(Ptr)->getOperand(0);
1495 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1496 if (GA->mayBeOverridden())
1498 Ptr = GA->getAliasee();
1502 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1503 } while (Visited.insert(Ptr).second);
1507 Int8Ptr = IRB.CreateBitCast(
1508 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1509 NamePrefix + "sroa_raw_cast");
1510 Int8PtrOffset = Offset;
1513 OffsetPtr = Int8PtrOffset == 0
1515 : IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1516 NamePrefix + "sroa_raw_idx");
1520 // On the off chance we were targeting i8*, guard the bitcast here.
1521 if (Ptr->getType() != PointerTy)
1522 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1527 /// \brief Test whether we can convert a value from the old to the new type.
1529 /// This predicate should be used to guard calls to convertValue in order to
1530 /// ensure that we only try to convert viable values. The strategy is that we
1531 /// will peel off single element struct and array wrappings to get to an
1532 /// underlying value, and convert that value.
1533 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1536 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1537 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1538 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1540 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1542 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1545 // We can convert pointers to integers and vice-versa. Same for vectors
1546 // of pointers and integers.
1547 OldTy = OldTy->getScalarType();
1548 NewTy = NewTy->getScalarType();
1549 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1550 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1552 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1560 /// \brief Generic routine to convert an SSA value to a value of a different
1563 /// This will try various different casting techniques, such as bitcasts,
1564 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1565 /// two types for viability with this routine.
1566 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1568 Type *OldTy = V->getType();
1569 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1574 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1575 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1576 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1577 return IRB.CreateZExt(V, NewITy);
1579 // See if we need inttoptr for this type pair. A cast involving both scalars
1580 // and vectors requires and additional bitcast.
1581 if (OldTy->getScalarType()->isIntegerTy() &&
1582 NewTy->getScalarType()->isPointerTy()) {
1583 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1584 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1585 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1588 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1589 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1590 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1593 return IRB.CreateIntToPtr(V, NewTy);
1596 // See if we need ptrtoint for this type pair. A cast involving both scalars
1597 // and vectors requires and additional bitcast.
1598 if (OldTy->getScalarType()->isPointerTy() &&
1599 NewTy->getScalarType()->isIntegerTy()) {
1600 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1601 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1602 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1605 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1606 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1607 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1610 return IRB.CreatePtrToInt(V, NewTy);
1613 return IRB.CreateBitCast(V, NewTy);
1616 /// \brief Test whether the given slice use can be promoted to a vector.
1618 /// This function is called to test each entry in a partioning which is slated
1619 /// for a single slice.
1621 isVectorPromotionViableForSlice(const DataLayout &DL, uint64_t SliceBeginOffset,
1622 uint64_t SliceEndOffset, VectorType *Ty,
1623 uint64_t ElementSize, const Slice &S) {
1624 // First validate the slice offsets.
1625 uint64_t BeginOffset =
1626 std::max(S.beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1627 uint64_t BeginIndex = BeginOffset / ElementSize;
1628 if (BeginIndex * ElementSize != BeginOffset ||
1629 BeginIndex >= Ty->getNumElements())
1631 uint64_t EndOffset =
1632 std::min(S.endOffset(), SliceEndOffset) - SliceBeginOffset;
1633 uint64_t EndIndex = EndOffset / ElementSize;
1634 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1637 assert(EndIndex > BeginIndex && "Empty vector!");
1638 uint64_t NumElements = EndIndex - BeginIndex;
1639 Type *SliceTy = (NumElements == 1)
1640 ? Ty->getElementType()
1641 : VectorType::get(Ty->getElementType(), NumElements);
1644 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1646 Use *U = S.getUse();
1648 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1649 if (MI->isVolatile())
1651 if (!S.isSplittable())
1652 return false; // Skip any unsplittable intrinsics.
1653 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1654 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1655 II->getIntrinsicID() != Intrinsic::lifetime_end)
1657 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1658 // Disable vector promotion when there are loads or stores of an FCA.
1660 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1661 if (LI->isVolatile())
1663 Type *LTy = LI->getType();
1664 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1665 assert(LTy->isIntegerTy());
1668 if (!canConvertValue(DL, SliceTy, LTy))
1670 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1671 if (SI->isVolatile())
1673 Type *STy = SI->getValueOperand()->getType();
1674 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1675 assert(STy->isIntegerTy());
1678 if (!canConvertValue(DL, STy, SliceTy))
1687 /// \brief Test whether the given alloca partitioning and range of slices can be
1688 /// promoted to a vector.
1690 /// This is a quick test to check whether we can rewrite a particular alloca
1691 /// partition (and its newly formed alloca) into a vector alloca with only
1692 /// whole-vector loads and stores such that it could be promoted to a vector
1693 /// SSA value. We only can ensure this for a limited set of operations, and we
1694 /// don't want to do the rewrites unless we are confident that the result will
1695 /// be promotable, so we have an early test here.
1697 isVectorPromotionViable(const DataLayout &DL, uint64_t SliceBeginOffset,
1698 uint64_t SliceEndOffset,
1699 AllocaSlices::const_range Slices,
1700 ArrayRef<AllocaSlices::iterator> SplitUses) {
1701 // Collect the candidate types for vector-based promotion. Also track whether
1702 // we have different element types.
1703 SmallVector<VectorType *, 4> CandidateTys;
1704 Type *CommonEltTy = nullptr;
1705 bool HaveCommonEltTy = true;
1706 auto CheckCandidateType = [&](Type *Ty) {
1707 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1708 CandidateTys.push_back(VTy);
1710 CommonEltTy = VTy->getElementType();
1711 else if (CommonEltTy != VTy->getElementType())
1712 HaveCommonEltTy = false;
1715 // Consider any loads or stores that are the exact size of the slice.
1716 for (const auto &S : Slices)
1717 if (S.beginOffset() == SliceBeginOffset &&
1718 S.endOffset() == SliceEndOffset) {
1719 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1720 CheckCandidateType(LI->getType());
1721 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1722 CheckCandidateType(SI->getValueOperand()->getType());
1725 // If we didn't find a vector type, nothing to do here.
1726 if (CandidateTys.empty())
1729 // Remove non-integer vector types if we had multiple common element types.
1730 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1731 // do that until all the backends are known to produce good code for all
1732 // integer vector types.
1733 if (!HaveCommonEltTy) {
1734 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
1735 [](VectorType *VTy) {
1736 return !VTy->getElementType()->isIntegerTy();
1738 CandidateTys.end());
1740 // If there were no integer vector types, give up.
1741 if (CandidateTys.empty())
1744 // Rank the remaining candidate vector types. This is easy because we know
1745 // they're all integer vectors. We sort by ascending number of elements.
1746 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1747 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1748 "Cannot have vector types of different sizes!");
1749 assert(RHSTy->getElementType()->isIntegerTy() &&
1750 "All non-integer types eliminated!");
1751 assert(LHSTy->getElementType()->isIntegerTy() &&
1752 "All non-integer types eliminated!");
1753 return RHSTy->getNumElements() < LHSTy->getNumElements();
1755 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1757 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1758 CandidateTys.end());
1760 // The only way to have the same element type in every vector type is to
1761 // have the same vector type. Check that and remove all but one.
1763 for (VectorType *VTy : CandidateTys) {
1764 assert(VTy->getElementType() == CommonEltTy &&
1765 "Unaccounted for element type!");
1766 assert(VTy == CandidateTys[0] &&
1767 "Different vector types with the same element type!");
1770 CandidateTys.resize(1);
1773 // Try each vector type, and return the one which works.
1774 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1775 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1777 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1778 // that aren't byte sized.
1779 if (ElementSize % 8)
1781 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1782 "vector size not a multiple of element size?");
1785 for (const auto &S : Slices)
1786 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1787 VTy, ElementSize, S))
1790 for (const auto &SI : SplitUses)
1791 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1792 VTy, ElementSize, *SI))
1797 for (VectorType *VTy : CandidateTys)
1798 if (CheckVectorTypeForPromotion(VTy))
1804 /// \brief Test whether a slice of an alloca is valid for integer widening.
1806 /// This implements the necessary checking for the \c isIntegerWideningViable
1807 /// test below on a single slice of the alloca.
1808 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1810 uint64_t AllocBeginOffset,
1811 uint64_t Size, const Slice &S,
1812 bool &WholeAllocaOp) {
1813 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1814 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1816 // We can't reasonably handle cases where the load or store extends past
1817 // the end of the aloca's type and into its padding.
1821 Use *U = S.getUse();
1823 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1824 if (LI->isVolatile())
1826 // Note that we don't count vector loads or stores as whole-alloca
1827 // operations which enable integer widening because we would prefer to use
1828 // vector widening instead.
1829 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1830 WholeAllocaOp = true;
1831 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1832 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1834 } else if (RelBegin != 0 || RelEnd != Size ||
1835 !canConvertValue(DL, AllocaTy, LI->getType())) {
1836 // Non-integer loads need to be convertible from the alloca type so that
1837 // they are promotable.
1840 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1841 Type *ValueTy = SI->getValueOperand()->getType();
1842 if (SI->isVolatile())
1844 // Note that we don't count vector loads or stores as whole-alloca
1845 // operations which enable integer widening because we would prefer to use
1846 // vector widening instead.
1847 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1848 WholeAllocaOp = true;
1849 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1850 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1852 } else if (RelBegin != 0 || RelEnd != Size ||
1853 !canConvertValue(DL, ValueTy, AllocaTy)) {
1854 // Non-integer stores need to be convertible to the alloca type so that
1855 // they are promotable.
1858 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1859 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1861 if (!S.isSplittable())
1862 return false; // Skip any unsplittable intrinsics.
1863 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1864 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1865 II->getIntrinsicID() != Intrinsic::lifetime_end)
1874 /// \brief Test whether the given alloca partition's integer operations can be
1875 /// widened to promotable ones.
1877 /// This is a quick test to check whether we can rewrite the integer loads and
1878 /// stores to a particular alloca into wider loads and stores and be able to
1879 /// promote the resulting alloca.
1881 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1882 uint64_t AllocBeginOffset,
1883 AllocaSlices::const_range Slices,
1884 ArrayRef<AllocaSlices::iterator> SplitUses) {
1885 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1886 // Don't create integer types larger than the maximum bitwidth.
1887 if (SizeInBits > IntegerType::MAX_INT_BITS)
1890 // Don't try to handle allocas with bit-padding.
1891 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1894 // We need to ensure that an integer type with the appropriate bitwidth can
1895 // be converted to the alloca type, whatever that is. We don't want to force
1896 // the alloca itself to have an integer type if there is a more suitable one.
1897 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1898 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1899 !canConvertValue(DL, IntTy, AllocaTy))
1902 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1904 // While examining uses, we ensure that the alloca has a covering load or
1905 // store. We don't want to widen the integer operations only to fail to
1906 // promote due to some other unsplittable entry (which we may make splittable
1907 // later). However, if there are only splittable uses, go ahead and assume
1908 // that we cover the alloca.
1909 bool WholeAllocaOp =
1910 Slices.begin() != Slices.end() ? false : DL.isLegalInteger(SizeInBits);
1912 for (const auto &S : Slices)
1913 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1917 for (const auto &SI : SplitUses)
1918 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1919 *SI, WholeAllocaOp))
1922 return WholeAllocaOp;
1925 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1926 IntegerType *Ty, uint64_t Offset,
1927 const Twine &Name) {
1928 DEBUG(dbgs() << " start: " << *V << "\n");
1929 IntegerType *IntTy = cast<IntegerType>(V->getType());
1930 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1931 "Element extends past full value");
1932 uint64_t ShAmt = 8 * Offset;
1933 if (DL.isBigEndian())
1934 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1936 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1937 DEBUG(dbgs() << " shifted: " << *V << "\n");
1939 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1940 "Cannot extract to a larger integer!");
1942 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1943 DEBUG(dbgs() << " trunced: " << *V << "\n");
1948 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1949 Value *V, uint64_t Offset, const Twine &Name) {
1950 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1951 IntegerType *Ty = cast<IntegerType>(V->getType());
1952 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1953 "Cannot insert a larger integer!");
1954 DEBUG(dbgs() << " start: " << *V << "\n");
1956 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1957 DEBUG(dbgs() << " extended: " << *V << "\n");
1959 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1960 "Element store outside of alloca store");
1961 uint64_t ShAmt = 8 * Offset;
1962 if (DL.isBigEndian())
1963 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1965 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1966 DEBUG(dbgs() << " shifted: " << *V << "\n");
1969 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1970 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1971 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1972 DEBUG(dbgs() << " masked: " << *Old << "\n");
1973 V = IRB.CreateOr(Old, V, Name + ".insert");
1974 DEBUG(dbgs() << " inserted: " << *V << "\n");
1979 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
1980 unsigned EndIndex, const Twine &Name) {
1981 VectorType *VecTy = cast<VectorType>(V->getType());
1982 unsigned NumElements = EndIndex - BeginIndex;
1983 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1985 if (NumElements == VecTy->getNumElements())
1988 if (NumElements == 1) {
1989 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1991 DEBUG(dbgs() << " extract: " << *V << "\n");
1995 SmallVector<Constant *, 8> Mask;
1996 Mask.reserve(NumElements);
1997 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1998 Mask.push_back(IRB.getInt32(i));
1999 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2000 ConstantVector::get(Mask), Name + ".extract");
2001 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2005 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2006 unsigned BeginIndex, const Twine &Name) {
2007 VectorType *VecTy = cast<VectorType>(Old->getType());
2008 assert(VecTy && "Can only insert a vector into a vector");
2010 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2012 // Single element to insert.
2013 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2015 DEBUG(dbgs() << " insert: " << *V << "\n");
2019 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2020 "Too many elements!");
2021 if (Ty->getNumElements() == VecTy->getNumElements()) {
2022 assert(V->getType() == VecTy && "Vector type mismatch");
2025 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2027 // When inserting a smaller vector into the larger to store, we first
2028 // use a shuffle vector to widen it with undef elements, and then
2029 // a second shuffle vector to select between the loaded vector and the
2031 SmallVector<Constant *, 8> Mask;
2032 Mask.reserve(VecTy->getNumElements());
2033 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2034 if (i >= BeginIndex && i < EndIndex)
2035 Mask.push_back(IRB.getInt32(i - BeginIndex));
2037 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2038 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2039 ConstantVector::get(Mask), Name + ".expand");
2040 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2043 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2044 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2046 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2048 DEBUG(dbgs() << " blend: " << *V << "\n");
2053 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2054 /// to use a new alloca.
2056 /// Also implements the rewriting to vector-based accesses when the partition
2057 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2059 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2060 // Befriend the base class so it can delegate to private visit methods.
2061 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2062 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2064 const DataLayout &DL;
2067 AllocaInst &OldAI, &NewAI;
2068 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2071 // This is a convenience and flag variable that will be null unless the new
2072 // alloca's integer operations should be widened to this integer type due to
2073 // passing isIntegerWideningViable above. If it is non-null, the desired
2074 // integer type will be stored here for easy access during rewriting.
2077 // If we are rewriting an alloca partition which can be written as pure
2078 // vector operations, we stash extra information here. When VecTy is
2079 // non-null, we have some strict guarantees about the rewritten alloca:
2080 // - The new alloca is exactly the size of the vector type here.
2081 // - The accesses all either map to the entire vector or to a single
2083 // - The set of accessing instructions is only one of those handled above
2084 // in isVectorPromotionViable. Generally these are the same access kinds
2085 // which are promotable via mem2reg.
2088 uint64_t ElementSize;
2090 // The original offset of the slice currently being rewritten relative to
2091 // the original alloca.
2092 uint64_t BeginOffset, EndOffset;
2093 // The new offsets of the slice currently being rewritten relative to the
2095 uint64_t NewBeginOffset, NewEndOffset;
2101 Instruction *OldPtr;
2103 // Track post-rewrite users which are PHI nodes and Selects.
2104 SmallPtrSetImpl<PHINode *> &PHIUsers;
2105 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2107 // Utility IR builder, whose name prefix is setup for each visited use, and
2108 // the insertion point is set to point to the user.
2112 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2113 AllocaInst &OldAI, AllocaInst &NewAI,
2114 uint64_t NewAllocaBeginOffset,
2115 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2116 VectorType *PromotableVecTy,
2117 SmallPtrSetImpl<PHINode *> &PHIUsers,
2118 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2119 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2120 NewAllocaBeginOffset(NewAllocaBeginOffset),
2121 NewAllocaEndOffset(NewAllocaEndOffset),
2122 NewAllocaTy(NewAI.getAllocatedType()),
2123 IntTy(IsIntegerPromotable
2126 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2128 VecTy(PromotableVecTy),
2129 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2130 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2131 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2132 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2133 IRB(NewAI.getContext(), ConstantFolder()) {
2135 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2136 "Only multiple-of-8 sized vector elements are viable");
2139 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2142 bool visit(AllocaSlices::const_iterator I) {
2143 bool CanSROA = true;
2144 BeginOffset = I->beginOffset();
2145 EndOffset = I->endOffset();
2146 IsSplittable = I->isSplittable();
2148 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2150 // Compute the intersecting offset range.
2151 assert(BeginOffset < NewAllocaEndOffset);
2152 assert(EndOffset > NewAllocaBeginOffset);
2153 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2154 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2156 SliceSize = NewEndOffset - NewBeginOffset;
2158 OldUse = I->getUse();
2159 OldPtr = cast<Instruction>(OldUse->get());
2161 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2162 IRB.SetInsertPoint(OldUserI);
2163 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2164 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2166 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2173 // Make sure the other visit overloads are visible.
2176 // Every instruction which can end up as a user must have a rewrite rule.
2177 bool visitInstruction(Instruction &I) {
2178 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2179 llvm_unreachable("No rewrite rule for this instruction!");
2182 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2183 // Note that the offset computation can use BeginOffset or NewBeginOffset
2184 // interchangeably for unsplit slices.
2185 assert(IsSplit || BeginOffset == NewBeginOffset);
2186 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2189 StringRef OldName = OldPtr->getName();
2190 // Skip through the last '.sroa.' component of the name.
2191 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2192 if (LastSROAPrefix != StringRef::npos) {
2193 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2194 // Look for an SROA slice index.
2195 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2196 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2197 // Strip the index and look for the offset.
2198 OldName = OldName.substr(IndexEnd + 1);
2199 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2200 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2201 // Strip the offset.
2202 OldName = OldName.substr(OffsetEnd + 1);
2205 // Strip any SROA suffixes as well.
2206 OldName = OldName.substr(0, OldName.find(".sroa_"));
2209 return getAdjustedPtr(IRB, DL, &NewAI,
2210 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2212 Twine(OldName) + "."
2219 /// \brief Compute suitable alignment to access this slice of the *new*
2222 /// You can optionally pass a type to this routine and if that type's ABI
2223 /// alignment is itself suitable, this will return zero.
2224 unsigned getSliceAlign(Type *Ty = nullptr) {
2225 unsigned NewAIAlign = NewAI.getAlignment();
2227 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2229 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2230 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2233 unsigned getIndex(uint64_t Offset) {
2234 assert(VecTy && "Can only call getIndex when rewriting a vector");
2235 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2236 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2237 uint32_t Index = RelOffset / ElementSize;
2238 assert(Index * ElementSize == RelOffset);
2242 void deleteIfTriviallyDead(Value *V) {
2243 Instruction *I = cast<Instruction>(V);
2244 if (isInstructionTriviallyDead(I))
2245 Pass.DeadInsts.insert(I);
2248 Value *rewriteVectorizedLoadInst() {
2249 unsigned BeginIndex = getIndex(NewBeginOffset);
2250 unsigned EndIndex = getIndex(NewEndOffset);
2251 assert(EndIndex > BeginIndex && "Empty vector!");
2253 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2254 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2257 Value *rewriteIntegerLoad(LoadInst &LI) {
2258 assert(IntTy && "We cannot insert an integer to the alloca");
2259 assert(!LI.isVolatile());
2260 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2261 V = convertValue(DL, IRB, V, IntTy);
2262 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2263 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2264 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2265 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2270 bool visitLoadInst(LoadInst &LI) {
2271 DEBUG(dbgs() << " original: " << LI << "\n");
2272 Value *OldOp = LI.getOperand(0);
2273 assert(OldOp == OldPtr);
2275 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2277 bool IsPtrAdjusted = false;
2280 V = rewriteVectorizedLoadInst();
2281 } else if (IntTy && LI.getType()->isIntegerTy()) {
2282 V = rewriteIntegerLoad(LI);
2283 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2284 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2285 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), LI.isVolatile(),
2288 Type *LTy = TargetTy->getPointerTo();
2289 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2290 getSliceAlign(TargetTy), LI.isVolatile(),
2292 IsPtrAdjusted = true;
2294 V = convertValue(DL, IRB, V, TargetTy);
2297 assert(!LI.isVolatile());
2298 assert(LI.getType()->isIntegerTy() &&
2299 "Only integer type loads and stores are split");
2300 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2301 "Split load isn't smaller than original load");
2302 assert(LI.getType()->getIntegerBitWidth() ==
2303 DL.getTypeStoreSizeInBits(LI.getType()) &&
2304 "Non-byte-multiple bit width");
2305 // Move the insertion point just past the load so that we can refer to it.
2306 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2307 // Create a placeholder value with the same type as LI to use as the
2308 // basis for the new value. This allows us to replace the uses of LI with
2309 // the computed value, and then replace the placeholder with LI, leaving
2310 // LI only used for this computation.
2311 Value *Placeholder =
2312 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2313 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset, "insert");
2314 LI.replaceAllUsesWith(V);
2315 Placeholder->replaceAllUsesWith(&LI);
2318 LI.replaceAllUsesWith(V);
2321 Pass.DeadInsts.insert(&LI);
2322 deleteIfTriviallyDead(OldOp);
2323 DEBUG(dbgs() << " to: " << *V << "\n");
2324 return !LI.isVolatile() && !IsPtrAdjusted;
2327 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2328 if (V->getType() != VecTy) {
2329 unsigned BeginIndex = getIndex(NewBeginOffset);
2330 unsigned EndIndex = getIndex(NewEndOffset);
2331 assert(EndIndex > BeginIndex && "Empty vector!");
2332 unsigned NumElements = EndIndex - BeginIndex;
2333 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2334 Type *SliceTy = (NumElements == 1)
2336 : VectorType::get(ElementTy, NumElements);
2337 if (V->getType() != SliceTy)
2338 V = convertValue(DL, IRB, V, SliceTy);
2340 // Mix in the existing elements.
2341 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2342 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2344 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2345 Pass.DeadInsts.insert(&SI);
2348 DEBUG(dbgs() << " to: " << *Store << "\n");
2352 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2353 assert(IntTy && "We cannot extract an integer from the alloca");
2354 assert(!SI.isVolatile());
2355 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2357 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2358 Old = convertValue(DL, IRB, Old, IntTy);
2359 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2360 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2361 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2363 V = convertValue(DL, IRB, V, NewAllocaTy);
2364 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2365 Pass.DeadInsts.insert(&SI);
2367 DEBUG(dbgs() << " to: " << *Store << "\n");
2371 bool visitStoreInst(StoreInst &SI) {
2372 DEBUG(dbgs() << " original: " << SI << "\n");
2373 Value *OldOp = SI.getOperand(1);
2374 assert(OldOp == OldPtr);
2376 Value *V = SI.getValueOperand();
2378 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2379 // alloca that should be re-examined after promoting this alloca.
2380 if (V->getType()->isPointerTy())
2381 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2382 Pass.PostPromotionWorklist.insert(AI);
2384 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2385 assert(!SI.isVolatile());
2386 assert(V->getType()->isIntegerTy() &&
2387 "Only integer type loads and stores are split");
2388 assert(V->getType()->getIntegerBitWidth() ==
2389 DL.getTypeStoreSizeInBits(V->getType()) &&
2390 "Non-byte-multiple bit width");
2391 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2392 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset, "extract");
2396 return rewriteVectorizedStoreInst(V, SI, OldOp);
2397 if (IntTy && V->getType()->isIntegerTy())
2398 return rewriteIntegerStore(V, SI);
2401 if (NewBeginOffset == NewAllocaBeginOffset &&
2402 NewEndOffset == NewAllocaEndOffset &&
2403 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2404 V = convertValue(DL, IRB, V, NewAllocaTy);
2405 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2408 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2409 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2413 Pass.DeadInsts.insert(&SI);
2414 deleteIfTriviallyDead(OldOp);
2416 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2417 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2420 /// \brief Compute an integer value from splatting an i8 across the given
2421 /// number of bytes.
2423 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2424 /// call this routine.
2425 /// FIXME: Heed the advice above.
2427 /// \param V The i8 value to splat.
2428 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2429 Value *getIntegerSplat(Value *V, unsigned Size) {
2430 assert(Size > 0 && "Expected a positive number of bytes.");
2431 IntegerType *VTy = cast<IntegerType>(V->getType());
2432 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2436 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2438 IRB.CreateZExt(V, SplatIntTy, "zext"),
2439 ConstantExpr::getUDiv(
2440 Constant::getAllOnesValue(SplatIntTy),
2441 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2447 /// \brief Compute a vector splat for a given element value.
2448 Value *getVectorSplat(Value *V, unsigned NumElements) {
2449 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2450 DEBUG(dbgs() << " splat: " << *V << "\n");
2454 bool visitMemSetInst(MemSetInst &II) {
2455 DEBUG(dbgs() << " original: " << II << "\n");
2456 assert(II.getRawDest() == OldPtr);
2458 // If the memset has a variable size, it cannot be split, just adjust the
2459 // pointer to the new alloca.
2460 if (!isa<Constant>(II.getLength())) {
2462 assert(NewBeginOffset == BeginOffset);
2463 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2464 Type *CstTy = II.getAlignmentCst()->getType();
2465 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2467 deleteIfTriviallyDead(OldPtr);
2471 // Record this instruction for deletion.
2472 Pass.DeadInsts.insert(&II);
2474 Type *AllocaTy = NewAI.getAllocatedType();
2475 Type *ScalarTy = AllocaTy->getScalarType();
2477 // If this doesn't map cleanly onto the alloca type, and that type isn't
2478 // a single value type, just emit a memset.
2479 if (!VecTy && !IntTy &&
2480 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2481 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2482 !AllocaTy->isSingleValueType() ||
2483 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2484 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2485 Type *SizeTy = II.getLength()->getType();
2486 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2487 CallInst *New = IRB.CreateMemSet(
2488 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2489 getSliceAlign(), II.isVolatile());
2491 DEBUG(dbgs() << " to: " << *New << "\n");
2495 // If we can represent this as a simple value, we have to build the actual
2496 // value to store, which requires expanding the byte present in memset to
2497 // a sensible representation for the alloca type. This is essentially
2498 // splatting the byte to a sufficiently wide integer, splatting it across
2499 // any desired vector width, and bitcasting to the final type.
2503 // If this is a memset of a vectorized alloca, insert it.
2504 assert(ElementTy == ScalarTy);
2506 unsigned BeginIndex = getIndex(NewBeginOffset);
2507 unsigned EndIndex = getIndex(NewEndOffset);
2508 assert(EndIndex > BeginIndex && "Empty vector!");
2509 unsigned NumElements = EndIndex - BeginIndex;
2510 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2513 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2514 Splat = convertValue(DL, IRB, Splat, ElementTy);
2515 if (NumElements > 1)
2516 Splat = getVectorSplat(Splat, NumElements);
2519 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2520 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2522 // If this is a memset on an alloca where we can widen stores, insert the
2524 assert(!II.isVolatile());
2526 uint64_t Size = NewEndOffset - NewBeginOffset;
2527 V = getIntegerSplat(II.getValue(), Size);
2529 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2530 EndOffset != NewAllocaBeginOffset)) {
2532 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2533 Old = convertValue(DL, IRB, Old, IntTy);
2534 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2535 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2537 assert(V->getType() == IntTy &&
2538 "Wrong type for an alloca wide integer!");
2540 V = convertValue(DL, IRB, V, AllocaTy);
2542 // Established these invariants above.
2543 assert(NewBeginOffset == NewAllocaBeginOffset);
2544 assert(NewEndOffset == NewAllocaEndOffset);
2546 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2547 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2548 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2550 V = convertValue(DL, IRB, V, AllocaTy);
2553 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2556 DEBUG(dbgs() << " to: " << *New << "\n");
2557 return !II.isVolatile();
2560 bool visitMemTransferInst(MemTransferInst &II) {
2561 // Rewriting of memory transfer instructions can be a bit tricky. We break
2562 // them into two categories: split intrinsics and unsplit intrinsics.
2564 DEBUG(dbgs() << " original: " << II << "\n");
2566 bool IsDest = &II.getRawDestUse() == OldUse;
2567 assert((IsDest && II.getRawDest() == OldPtr) ||
2568 (!IsDest && II.getRawSource() == OldPtr));
2570 unsigned SliceAlign = getSliceAlign();
2572 // For unsplit intrinsics, we simply modify the source and destination
2573 // pointers in place. This isn't just an optimization, it is a matter of
2574 // correctness. With unsplit intrinsics we may be dealing with transfers
2575 // within a single alloca before SROA ran, or with transfers that have
2576 // a variable length. We may also be dealing with memmove instead of
2577 // memcpy, and so simply updating the pointers is the necessary for us to
2578 // update both source and dest of a single call.
2579 if (!IsSplittable) {
2580 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2582 II.setDest(AdjustedPtr);
2584 II.setSource(AdjustedPtr);
2586 if (II.getAlignment() > SliceAlign) {
2587 Type *CstTy = II.getAlignmentCst()->getType();
2589 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2592 DEBUG(dbgs() << " to: " << II << "\n");
2593 deleteIfTriviallyDead(OldPtr);
2596 // For split transfer intrinsics we have an incredibly useful assurance:
2597 // the source and destination do not reside within the same alloca, and at
2598 // least one of them does not escape. This means that we can replace
2599 // memmove with memcpy, and we don't need to worry about all manner of
2600 // downsides to splitting and transforming the operations.
2602 // If this doesn't map cleanly onto the alloca type, and that type isn't
2603 // a single value type, just emit a memcpy.
2606 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2607 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2608 !NewAI.getAllocatedType()->isSingleValueType());
2610 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2611 // size hasn't been shrunk based on analysis of the viable range, this is
2613 if (EmitMemCpy && &OldAI == &NewAI) {
2614 // Ensure the start lines up.
2615 assert(NewBeginOffset == BeginOffset);
2617 // Rewrite the size as needed.
2618 if (NewEndOffset != EndOffset)
2619 II.setLength(ConstantInt::get(II.getLength()->getType(),
2620 NewEndOffset - NewBeginOffset));
2623 // Record this instruction for deletion.
2624 Pass.DeadInsts.insert(&II);
2626 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2627 // alloca that should be re-examined after rewriting this instruction.
2628 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2629 if (AllocaInst *AI =
2630 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2631 assert(AI != &OldAI && AI != &NewAI &&
2632 "Splittable transfers cannot reach the same alloca on both ends.");
2633 Pass.Worklist.insert(AI);
2636 Type *OtherPtrTy = OtherPtr->getType();
2637 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2639 // Compute the relative offset for the other pointer within the transfer.
2640 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2641 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2642 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2643 OtherOffset.zextOrTrunc(64).getZExtValue());
2646 // Compute the other pointer, folding as much as possible to produce
2647 // a single, simple GEP in most cases.
2648 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2649 OtherPtr->getName() + ".");
2651 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2652 Type *SizeTy = II.getLength()->getType();
2653 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2655 CallInst *New = IRB.CreateMemCpy(
2656 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2657 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2659 DEBUG(dbgs() << " to: " << *New << "\n");
2663 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2664 NewEndOffset == NewAllocaEndOffset;
2665 uint64_t Size = NewEndOffset - NewBeginOffset;
2666 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2667 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2668 unsigned NumElements = EndIndex - BeginIndex;
2669 IntegerType *SubIntTy =
2670 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2672 // Reset the other pointer type to match the register type we're going to
2673 // use, but using the address space of the original other pointer.
2674 if (VecTy && !IsWholeAlloca) {
2675 if (NumElements == 1)
2676 OtherPtrTy = VecTy->getElementType();
2678 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2680 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2681 } else if (IntTy && !IsWholeAlloca) {
2682 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2684 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2687 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2688 OtherPtr->getName() + ".");
2689 unsigned SrcAlign = OtherAlign;
2690 Value *DstPtr = &NewAI;
2691 unsigned DstAlign = SliceAlign;
2693 std::swap(SrcPtr, DstPtr);
2694 std::swap(SrcAlign, DstAlign);
2698 if (VecTy && !IsWholeAlloca && !IsDest) {
2699 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2700 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2701 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2702 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2703 Src = convertValue(DL, IRB, Src, IntTy);
2704 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2705 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2708 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2711 if (VecTy && !IsWholeAlloca && IsDest) {
2713 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2714 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2715 } else if (IntTy && !IsWholeAlloca && IsDest) {
2717 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2718 Old = convertValue(DL, IRB, Old, IntTy);
2719 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2720 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2721 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2724 StoreInst *Store = cast<StoreInst>(
2725 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2727 DEBUG(dbgs() << " to: " << *Store << "\n");
2728 return !II.isVolatile();
2731 bool visitIntrinsicInst(IntrinsicInst &II) {
2732 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2733 II.getIntrinsicID() == Intrinsic::lifetime_end);
2734 DEBUG(dbgs() << " original: " << II << "\n");
2735 assert(II.getArgOperand(1) == OldPtr);
2737 // Record this instruction for deletion.
2738 Pass.DeadInsts.insert(&II);
2741 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2742 NewEndOffset - NewBeginOffset);
2743 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2745 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2746 New = IRB.CreateLifetimeStart(Ptr, Size);
2748 New = IRB.CreateLifetimeEnd(Ptr, Size);
2751 DEBUG(dbgs() << " to: " << *New << "\n");
2755 bool visitPHINode(PHINode &PN) {
2756 DEBUG(dbgs() << " original: " << PN << "\n");
2757 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2758 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2760 // We would like to compute a new pointer in only one place, but have it be
2761 // as local as possible to the PHI. To do that, we re-use the location of
2762 // the old pointer, which necessarily must be in the right position to
2763 // dominate the PHI.
2764 IRBuilderTy PtrBuilder(IRB);
2765 if (isa<PHINode>(OldPtr))
2766 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
2768 PtrBuilder.SetInsertPoint(OldPtr);
2769 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2771 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2772 // Replace the operands which were using the old pointer.
2773 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2775 DEBUG(dbgs() << " to: " << PN << "\n");
2776 deleteIfTriviallyDead(OldPtr);
2778 // PHIs can't be promoted on their own, but often can be speculated. We
2779 // check the speculation outside of the rewriter so that we see the
2780 // fully-rewritten alloca.
2781 PHIUsers.insert(&PN);
2785 bool visitSelectInst(SelectInst &SI) {
2786 DEBUG(dbgs() << " original: " << SI << "\n");
2787 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2788 "Pointer isn't an operand!");
2789 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2790 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2792 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2793 // Replace the operands which were using the old pointer.
2794 if (SI.getOperand(1) == OldPtr)
2795 SI.setOperand(1, NewPtr);
2796 if (SI.getOperand(2) == OldPtr)
2797 SI.setOperand(2, NewPtr);
2799 DEBUG(dbgs() << " to: " << SI << "\n");
2800 deleteIfTriviallyDead(OldPtr);
2802 // Selects can't be promoted on their own, but often can be speculated. We
2803 // check the speculation outside of the rewriter so that we see the
2804 // fully-rewritten alloca.
2805 SelectUsers.insert(&SI);
2812 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2814 /// This pass aggressively rewrites all aggregate loads and stores on
2815 /// a particular pointer (or any pointer derived from it which we can identify)
2816 /// with scalar loads and stores.
2817 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2818 // Befriend the base class so it can delegate to private visit methods.
2819 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2821 const DataLayout &DL;
2823 /// Queue of pointer uses to analyze and potentially rewrite.
2824 SmallVector<Use *, 8> Queue;
2826 /// Set to prevent us from cycling with phi nodes and loops.
2827 SmallPtrSet<User *, 8> Visited;
2829 /// The current pointer use being rewritten. This is used to dig up the used
2830 /// value (as opposed to the user).
2834 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2836 /// Rewrite loads and stores through a pointer and all pointers derived from
2838 bool rewrite(Instruction &I) {
2839 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2841 bool Changed = false;
2842 while (!Queue.empty()) {
2843 U = Queue.pop_back_val();
2844 Changed |= visit(cast<Instruction>(U->getUser()));
2850 /// Enqueue all the users of the given instruction for further processing.
2851 /// This uses a set to de-duplicate users.
2852 void enqueueUsers(Instruction &I) {
2853 for (Use &U : I.uses())
2854 if (Visited.insert(U.getUser()).second)
2855 Queue.push_back(&U);
2858 // Conservative default is to not rewrite anything.
2859 bool visitInstruction(Instruction &I) { return false; }
2861 /// \brief Generic recursive split emission class.
2862 template <typename Derived> class OpSplitter {
2864 /// The builder used to form new instructions.
2866 /// The indices which to be used with insert- or extractvalue to select the
2867 /// appropriate value within the aggregate.
2868 SmallVector<unsigned, 4> Indices;
2869 /// The indices to a GEP instruction which will move Ptr to the correct slot
2870 /// within the aggregate.
2871 SmallVector<Value *, 4> GEPIndices;
2872 /// The base pointer of the original op, used as a base for GEPing the
2873 /// split operations.
2876 /// Initialize the splitter with an insertion point, Ptr and start with a
2877 /// single zero GEP index.
2878 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2879 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2882 /// \brief Generic recursive split emission routine.
2884 /// This method recursively splits an aggregate op (load or store) into
2885 /// scalar or vector ops. It splits recursively until it hits a single value
2886 /// and emits that single value operation via the template argument.
2888 /// The logic of this routine relies on GEPs and insertvalue and
2889 /// extractvalue all operating with the same fundamental index list, merely
2890 /// formatted differently (GEPs need actual values).
2892 /// \param Ty The type being split recursively into smaller ops.
2893 /// \param Agg The aggregate value being built up or stored, depending on
2894 /// whether this is splitting a load or a store respectively.
2895 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2896 if (Ty->isSingleValueType())
2897 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2899 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2900 unsigned OldSize = Indices.size();
2902 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2904 assert(Indices.size() == OldSize && "Did not return to the old size");
2905 Indices.push_back(Idx);
2906 GEPIndices.push_back(IRB.getInt32(Idx));
2907 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2908 GEPIndices.pop_back();
2914 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2915 unsigned OldSize = Indices.size();
2917 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2919 assert(Indices.size() == OldSize && "Did not return to the old size");
2920 Indices.push_back(Idx);
2921 GEPIndices.push_back(IRB.getInt32(Idx));
2922 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2923 GEPIndices.pop_back();
2929 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2933 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2934 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2935 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2937 /// Emit a leaf load of a single value. This is called at the leaves of the
2938 /// recursive emission to actually load values.
2939 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2940 assert(Ty->isSingleValueType());
2941 // Load the single value and insert it using the indices.
2942 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2943 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2944 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2945 DEBUG(dbgs() << " to: " << *Load << "\n");
2949 bool visitLoadInst(LoadInst &LI) {
2950 assert(LI.getPointerOperand() == *U);
2951 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2954 // We have an aggregate being loaded, split it apart.
2955 DEBUG(dbgs() << " original: " << LI << "\n");
2956 LoadOpSplitter Splitter(&LI, *U);
2957 Value *V = UndefValue::get(LI.getType());
2958 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2959 LI.replaceAllUsesWith(V);
2960 LI.eraseFromParent();
2964 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2965 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2966 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2968 /// Emit a leaf store of a single value. This is called at the leaves of the
2969 /// recursive emission to actually produce stores.
2970 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2971 assert(Ty->isSingleValueType());
2972 // Extract the single value and store it using the indices.
2973 Value *Store = IRB.CreateStore(
2974 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2975 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2977 DEBUG(dbgs() << " to: " << *Store << "\n");
2981 bool visitStoreInst(StoreInst &SI) {
2982 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2984 Value *V = SI.getValueOperand();
2985 if (V->getType()->isSingleValueType())
2988 // We have an aggregate being stored, split it apart.
2989 DEBUG(dbgs() << " original: " << SI << "\n");
2990 StoreOpSplitter Splitter(&SI, *U);
2991 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2992 SI.eraseFromParent();
2996 bool visitBitCastInst(BitCastInst &BC) {
3001 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3006 bool visitPHINode(PHINode &PN) {
3011 bool visitSelectInst(SelectInst &SI) {
3018 /// \brief Strip aggregate type wrapping.
3020 /// This removes no-op aggregate types wrapping an underlying type. It will
3021 /// strip as many layers of types as it can without changing either the type
3022 /// size or the allocated size.
3023 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3024 if (Ty->isSingleValueType())
3027 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3028 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3031 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3032 InnerTy = ArrTy->getElementType();
3033 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3034 const StructLayout *SL = DL.getStructLayout(STy);
3035 unsigned Index = SL->getElementContainingOffset(0);
3036 InnerTy = STy->getElementType(Index);
3041 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3042 TypeSize > DL.getTypeSizeInBits(InnerTy))
3045 return stripAggregateTypeWrapping(DL, InnerTy);
3048 /// \brief Try to find a partition of the aggregate type passed in for a given
3049 /// offset and size.
3051 /// This recurses through the aggregate type and tries to compute a subtype
3052 /// based on the offset and size. When the offset and size span a sub-section
3053 /// of an array, it will even compute a new array type for that sub-section,
3054 /// and the same for structs.
3056 /// Note that this routine is very strict and tries to find a partition of the
3057 /// type which produces the *exact* right offset and size. It is not forgiving
3058 /// when the size or offset cause either end of type-based partition to be off.
3059 /// Also, this is a best-effort routine. It is reasonable to give up and not
3060 /// return a type if necessary.
3061 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3063 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3064 return stripAggregateTypeWrapping(DL, Ty);
3065 if (Offset > DL.getTypeAllocSize(Ty) ||
3066 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3069 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3070 // We can't partition pointers...
3071 if (SeqTy->isPointerTy())
3074 Type *ElementTy = SeqTy->getElementType();
3075 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3076 uint64_t NumSkippedElements = Offset / ElementSize;
3077 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3078 if (NumSkippedElements >= ArrTy->getNumElements())
3080 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3081 if (NumSkippedElements >= VecTy->getNumElements())
3084 Offset -= NumSkippedElements * ElementSize;
3086 // First check if we need to recurse.
3087 if (Offset > 0 || Size < ElementSize) {
3088 // Bail if the partition ends in a different array element.
3089 if ((Offset + Size) > ElementSize)
3091 // Recurse through the element type trying to peel off offset bytes.
3092 return getTypePartition(DL, ElementTy, Offset, Size);
3094 assert(Offset == 0);
3096 if (Size == ElementSize)
3097 return stripAggregateTypeWrapping(DL, ElementTy);
3098 assert(Size > ElementSize);
3099 uint64_t NumElements = Size / ElementSize;
3100 if (NumElements * ElementSize != Size)
3102 return ArrayType::get(ElementTy, NumElements);
3105 StructType *STy = dyn_cast<StructType>(Ty);
3109 const StructLayout *SL = DL.getStructLayout(STy);
3110 if (Offset >= SL->getSizeInBytes())
3112 uint64_t EndOffset = Offset + Size;
3113 if (EndOffset > SL->getSizeInBytes())
3116 unsigned Index = SL->getElementContainingOffset(Offset);
3117 Offset -= SL->getElementOffset(Index);
3119 Type *ElementTy = STy->getElementType(Index);
3120 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3121 if (Offset >= ElementSize)
3122 return nullptr; // The offset points into alignment padding.
3124 // See if any partition must be contained by the element.
3125 if (Offset > 0 || Size < ElementSize) {
3126 if ((Offset + Size) > ElementSize)
3128 return getTypePartition(DL, ElementTy, Offset, Size);
3130 assert(Offset == 0);
3132 if (Size == ElementSize)
3133 return stripAggregateTypeWrapping(DL, ElementTy);
3135 StructType::element_iterator EI = STy->element_begin() + Index,
3136 EE = STy->element_end();
3137 if (EndOffset < SL->getSizeInBytes()) {
3138 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3139 if (Index == EndIndex)
3140 return nullptr; // Within a single element and its padding.
3142 // Don't try to form "natural" types if the elements don't line up with the
3144 // FIXME: We could potentially recurse down through the last element in the
3145 // sub-struct to find a natural end point.
3146 if (SL->getElementOffset(EndIndex) != EndOffset)
3149 assert(Index < EndIndex);
3150 EE = STy->element_begin() + EndIndex;
3153 // Try to build up a sub-structure.
3155 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3156 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3157 if (Size != SubSL->getSizeInBytes())
3158 return nullptr; // The sub-struct doesn't have quite the size needed.
3163 /// \brief Rewrite an alloca partition's users.
3165 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3166 /// to rewrite uses of an alloca partition to be conducive for SSA value
3167 /// promotion. If the partition needs a new, more refined alloca, this will
3168 /// build that new alloca, preserving as much type information as possible, and
3169 /// rewrite the uses of the old alloca to point at the new one and have the
3170 /// appropriate new offsets. It also evaluates how successful the rewrite was
3171 /// at enabling promotion and if it was successful queues the alloca to be
3173 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3174 AllocaSlices::iterator B, AllocaSlices::iterator E,
3175 int64_t BeginOffset, int64_t EndOffset,
3176 ArrayRef<AllocaSlices::iterator> SplitUses) {
3177 assert(BeginOffset < EndOffset);
3178 uint64_t SliceSize = EndOffset - BeginOffset;
3180 // Try to compute a friendly type for this partition of the alloca. This
3181 // won't always succeed, in which case we fall back to a legal integer type
3182 // or an i8 array of an appropriate size.
3183 Type *SliceTy = nullptr;
3184 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3185 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3186 SliceTy = CommonUseTy;
3188 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3189 BeginOffset, SliceSize))
3190 SliceTy = TypePartitionTy;
3191 if ((!SliceTy || (SliceTy->isArrayTy() &&
3192 SliceTy->getArrayElementType()->isIntegerTy())) &&
3193 DL->isLegalInteger(SliceSize * 8))
3194 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3196 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3197 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3199 bool IsIntegerPromotable = isIntegerWideningViable(
3200 *DL, SliceTy, BeginOffset, AllocaSlices::const_range(B, E), SplitUses);
3205 : isVectorPromotionViable(*DL, BeginOffset, EndOffset,
3206 AllocaSlices::const_range(B, E), SplitUses);
3210 // Check for the case where we're going to rewrite to a new alloca of the
3211 // exact same type as the original, and with the same access offsets. In that
3212 // case, re-use the existing alloca, but still run through the rewriter to
3213 // perform phi and select speculation.
3215 if (SliceTy == AI.getAllocatedType()) {
3216 assert(BeginOffset == 0 && "Non-zero begin offset but same alloca type");
3218 // FIXME: We should be able to bail at this point with "nothing changed".
3219 // FIXME: We might want to defer PHI speculation until after here.
3221 unsigned Alignment = AI.getAlignment();
3223 // The minimum alignment which users can rely on when the explicit
3224 // alignment is omitted or zero is that required by the ABI for this
3226 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3228 Alignment = MinAlign(Alignment, BeginOffset);
3229 // If we will get at least this much alignment from the type alone, leave
3230 // the alloca's alignment unconstrained.
3231 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3234 new AllocaInst(SliceTy, nullptr, Alignment,
3235 AI.getName() + ".sroa." + Twine(B - AS.begin()), &AI);
3238 // Migrate debug information from the old alloca to the new alloca
3239 // and the individial slices.
3240 if (DbgDeclareInst *DbgDecl = DbgDeclares.lookup(&AI)) {
3241 DIVariable Var(DbgDecl->getVariable());
3243 DIBuilder DIB(*AI.getParent()->getParent()->getParent(),
3244 /*AllowUnresolved*/ false);
3245 // Create a piece expression describing the slice, if the new slize is
3246 // smaller than the old alloca or the old alloca already was described
3247 // with a piece. It would be even better to just compare against the size
3248 // of the type described in the debug info, but then we would need to
3249 // build an expensive DIRefMap.
3250 if (SliceSize < DL->getTypeAllocSize(AI.getAllocatedType()) ||
3251 DIExpression(DbgDecl->getExpression()).isVariablePiece())
3252 Piece = DIB.createPieceExpression(BeginOffset, SliceSize);
3253 Instruction *NewDDI = DIB.insertDeclare(NewAI, Var, Piece, &AI);
3254 NewDDI->setDebugLoc(DbgDecl->getDebugLoc());
3255 DbgDeclares.insert(std::make_pair(NewAI, cast<DbgDeclareInst>(NewDDI)));
3256 DeadInsts.insert(DbgDecl);
3260 DEBUG(dbgs() << "Rewriting alloca partition "
3261 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3264 // Track the high watermark on the worklist as it is only relevant for
3265 // promoted allocas. We will reset it to this point if the alloca is not in
3266 // fact scheduled for promotion.
3267 unsigned PPWOldSize = PostPromotionWorklist.size();
3268 unsigned NumUses = 0;
3269 SmallPtrSet<PHINode *, 8> PHIUsers;
3270 SmallPtrSet<SelectInst *, 8> SelectUsers;
3272 AllocaSliceRewriter Rewriter(*DL, AS, *this, AI, *NewAI, BeginOffset,
3273 EndOffset, IsIntegerPromotable, VecTy, PHIUsers,
3275 bool Promotable = true;
3276 for (auto &SplitUse : SplitUses) {
3277 DEBUG(dbgs() << " rewriting split ");
3278 DEBUG(AS.printSlice(dbgs(), SplitUse, ""));
3279 Promotable &= Rewriter.visit(SplitUse);
3282 for (AllocaSlices::iterator I = B; I != E; ++I) {
3283 DEBUG(dbgs() << " rewriting ");
3284 DEBUG(AS.printSlice(dbgs(), I, ""));
3285 Promotable &= Rewriter.visit(I);
3289 NumAllocaPartitionUses += NumUses;
3290 MaxUsesPerAllocaPartition =
3291 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3293 // Now that we've processed all the slices in the new partition, check if any
3294 // PHIs or Selects would block promotion.
3295 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3298 if (!isSafePHIToSpeculate(**I, DL)) {
3301 SelectUsers.clear();
3304 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3305 E = SelectUsers.end();
3307 if (!isSafeSelectToSpeculate(**I, DL)) {
3310 SelectUsers.clear();
3315 if (PHIUsers.empty() && SelectUsers.empty()) {
3316 // Promote the alloca.
3317 PromotableAllocas.push_back(NewAI);
3319 // If we have either PHIs or Selects to speculate, add them to those
3320 // worklists and re-queue the new alloca so that we promote in on the
3322 for (PHINode *PHIUser : PHIUsers)
3323 SpeculatablePHIs.insert(PHIUser);
3324 for (SelectInst *SelectUser : SelectUsers)
3325 SpeculatableSelects.insert(SelectUser);
3326 Worklist.insert(NewAI);
3329 // If we can't promote the alloca, iterate on it to check for new
3330 // refinements exposed by splitting the current alloca. Don't iterate on an
3331 // alloca which didn't actually change and didn't get promoted.
3333 Worklist.insert(NewAI);
3335 // Drop any post-promotion work items if promotion didn't happen.
3336 while (PostPromotionWorklist.size() > PPWOldSize)
3337 PostPromotionWorklist.pop_back();
3344 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3345 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3346 if (Offset >= MaxSplitUseEndOffset) {
3348 MaxSplitUseEndOffset = 0;
3352 size_t SplitUsesOldSize = SplitUses.size();
3353 SplitUses.erase(std::remove_if(
3354 SplitUses.begin(), SplitUses.end(),
3355 [Offset](const AllocaSlices::iterator &I) {
3356 return I->endOffset() <= Offset;
3359 if (SplitUsesOldSize == SplitUses.size())
3362 // Recompute the max. While this is linear, so is remove_if.
3363 MaxSplitUseEndOffset = 0;
3364 for (AllocaSlices::iterator SplitUse : SplitUses)
3365 MaxSplitUseEndOffset =
3366 std::max(SplitUse->endOffset(), MaxSplitUseEndOffset);
3369 /// \brief Walks the slices of an alloca and form partitions based on them,
3370 /// rewriting each of their uses.
3371 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3372 if (AS.begin() == AS.end())
3375 unsigned NumPartitions = 0;
3376 bool Changed = false;
3377 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3378 uint64_t MaxSplitUseEndOffset = 0;
3380 uint64_t BeginOffset = AS.begin()->beginOffset();
3382 for (AllocaSlices::iterator SI = AS.begin(), SJ = std::next(SI),
3384 SI != SE; SI = SJ) {
3385 uint64_t MaxEndOffset = SI->endOffset();
3387 if (!SI->isSplittable()) {
3388 // When we're forming an unsplittable region, it must always start at the
3389 // first slice and will extend through its end.
3390 assert(BeginOffset == SI->beginOffset());
3392 // Form a partition including all of the overlapping slices with this
3393 // unsplittable slice.
3394 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3395 if (!SJ->isSplittable())
3396 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3400 assert(SI->isSplittable()); // Established above.
3402 // Collect all of the overlapping splittable slices.
3403 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3404 SJ->isSplittable()) {
3405 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3409 // Back up MaxEndOffset and SJ if we ended the span early when
3410 // encountering an unsplittable slice.
3411 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3412 assert(!SJ->isSplittable());
3413 MaxEndOffset = SJ->beginOffset();
3417 // Check if we have managed to move the end offset forward yet. If so,
3418 // we'll have to rewrite uses and erase old split uses.
3419 if (BeginOffset < MaxEndOffset) {
3420 // Rewrite a sequence of overlapping slices.
3421 Changed |= rewritePartition(AI, AS, SI, SJ, BeginOffset, MaxEndOffset,
3425 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3428 // Accumulate all the splittable slices from the [SI,SJ) region which
3429 // overlap going forward.
3430 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3431 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3432 SplitUses.push_back(SK);
3433 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3436 // If we're already at the end and we have no split uses, we're done.
3437 if (SJ == SE && SplitUses.empty())
3440 // If we have no split uses or no gap in offsets, we're ready to move to
3442 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3443 BeginOffset = SJ->beginOffset();
3447 // Even if we have split slices, if the next slice is splittable and the
3448 // split slices reach it, we can simply set up the beginning offset of the
3449 // next iteration to bridge between them.
3450 if (SJ != SE && SJ->isSplittable() &&
3451 MaxSplitUseEndOffset > SJ->beginOffset()) {
3452 BeginOffset = MaxEndOffset;
3456 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3458 uint64_t PostSplitEndOffset =
3459 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3461 Changed |= rewritePartition(AI, AS, SJ, SJ, MaxEndOffset,
3462 PostSplitEndOffset, SplitUses);
3466 break; // Skip the rest, we don't need to do any cleanup.
3468 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3469 PostSplitEndOffset);
3471 // Now just reset the begin offset for the next iteration.
3472 BeginOffset = SJ->beginOffset();
3475 NumAllocaPartitions += NumPartitions;
3476 MaxPartitionsPerAlloca =
3477 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3482 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3483 void SROA::clobberUse(Use &U) {
3485 // Replace the use with an undef value.
3486 U = UndefValue::get(OldV->getType());
3488 // Check for this making an instruction dead. We have to garbage collect
3489 // all the dead instructions to ensure the uses of any alloca end up being
3491 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3492 if (isInstructionTriviallyDead(OldI)) {
3493 DeadInsts.insert(OldI);
3497 /// \brief Analyze an alloca for SROA.
3499 /// This analyzes the alloca to ensure we can reason about it, builds
3500 /// the slices of the alloca, and then hands it off to be split and
3501 /// rewritten as needed.
3502 bool SROA::runOnAlloca(AllocaInst &AI) {
3503 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3504 ++NumAllocasAnalyzed;
3506 // Special case dead allocas, as they're trivial.
3507 if (AI.use_empty()) {
3508 AI.eraseFromParent();
3512 // Skip alloca forms that this analysis can't handle.
3513 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3514 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3517 bool Changed = false;
3519 // First, split any FCA loads and stores touching this alloca to promote
3520 // better splitting and promotion opportunities.
3521 AggLoadStoreRewriter AggRewriter(*DL);
3522 Changed |= AggRewriter.rewrite(AI);
3524 // Build the slices using a recursive instruction-visiting builder.
3525 AllocaSlices AS(*DL, AI);
3526 DEBUG(AS.print(dbgs()));
3530 // Delete all the dead users of this alloca before splitting and rewriting it.
3531 for (Instruction *DeadUser : AS.getDeadUsers()) {
3532 // Free up everything used by this instruction.
3533 for (Use &DeadOp : DeadUser->operands())
3536 // Now replace the uses of this instruction.
3537 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
3539 // And mark it for deletion.
3540 DeadInsts.insert(DeadUser);
3543 for (Use *DeadOp : AS.getDeadOperands()) {
3544 clobberUse(*DeadOp);
3548 // No slices to split. Leave the dead alloca for a later pass to clean up.
3549 if (AS.begin() == AS.end())
3552 Changed |= splitAlloca(AI, AS);
3554 DEBUG(dbgs() << " Speculating PHIs\n");
3555 while (!SpeculatablePHIs.empty())
3556 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3558 DEBUG(dbgs() << " Speculating Selects\n");
3559 while (!SpeculatableSelects.empty())
3560 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3565 /// \brief Delete the dead instructions accumulated in this run.
3567 /// Recursively deletes the dead instructions we've accumulated. This is done
3568 /// at the very end to maximize locality of the recursive delete and to
3569 /// minimize the problems of invalidated instruction pointers as such pointers
3570 /// are used heavily in the intermediate stages of the algorithm.
3572 /// We also record the alloca instructions deleted here so that they aren't
3573 /// subsequently handed to mem2reg to promote.
3574 void SROA::deleteDeadInstructions(
3575 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
3576 while (!DeadInsts.empty()) {
3577 Instruction *I = DeadInsts.pop_back_val();
3578 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3580 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3582 for (Use &Operand : I->operands())
3583 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3584 // Zero out the operand and see if it becomes trivially dead.
3586 if (isInstructionTriviallyDead(U))
3587 DeadInsts.insert(U);
3590 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3591 DeletedAllocas.insert(AI);
3594 I->eraseFromParent();
3598 static void enqueueUsersInWorklist(Instruction &I,
3599 SmallVectorImpl<Instruction *> &Worklist,
3600 SmallPtrSetImpl<Instruction *> &Visited) {
3601 for (User *U : I.users())
3602 if (Visited.insert(cast<Instruction>(U)).second)
3603 Worklist.push_back(cast<Instruction>(U));
3606 /// \brief Promote the allocas, using the best available technique.
3608 /// This attempts to promote whatever allocas have been identified as viable in
3609 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3610 /// If there is a domtree available, we attempt to promote using the full power
3611 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3612 /// based on the SSAUpdater utilities. This function returns whether any
3613 /// promotion occurred.
3614 bool SROA::promoteAllocas(Function &F) {
3615 if (PromotableAllocas.empty())
3618 NumPromoted += PromotableAllocas.size();
3620 if (DT && !ForceSSAUpdater) {
3621 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3622 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AT);
3623 PromotableAllocas.clear();
3627 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3629 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
3630 SmallVector<Instruction *, 64> Insts;
3632 // We need a worklist to walk the uses of each alloca.
3633 SmallVector<Instruction *, 8> Worklist;
3634 SmallPtrSet<Instruction *, 8> Visited;
3635 SmallVector<Instruction *, 32> DeadInsts;
3637 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3638 AllocaInst *AI = PromotableAllocas[Idx];
3643 enqueueUsersInWorklist(*AI, Worklist, Visited);
3645 while (!Worklist.empty()) {
3646 Instruction *I = Worklist.pop_back_val();
3648 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3649 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3650 // leading to them) here. Eventually it should use them to optimize the
3651 // scalar values produced.
3652 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3653 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3654 II->getIntrinsicID() == Intrinsic::lifetime_end);
3655 II->eraseFromParent();
3659 // Push the loads and stores we find onto the list. SROA will already
3660 // have validated that all loads and stores are viable candidates for
3662 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3663 assert(LI->getType() == AI->getAllocatedType());
3664 Insts.push_back(LI);
3667 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3668 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3669 Insts.push_back(SI);
3673 // For everything else, we know that only no-op bitcasts and GEPs will
3674 // make it this far, just recurse through them and recall them for later
3676 DeadInsts.push_back(I);
3677 enqueueUsersInWorklist(*I, Worklist, Visited);
3679 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3680 while (!DeadInsts.empty())
3681 DeadInsts.pop_back_val()->eraseFromParent();
3682 AI->eraseFromParent();
3685 PromotableAllocas.clear();
3689 bool SROA::runOnFunction(Function &F) {
3690 if (skipOptnoneFunction(F))
3693 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3694 C = &F.getContext();
3695 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3697 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3700 DL = &DLP->getDataLayout();
3701 DominatorTreeWrapperPass *DTWP =
3702 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3703 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3704 AT = &getAnalysis<AssumptionTracker>();
3706 BasicBlock &EntryBB = F.getEntryBlock();
3707 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3709 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3710 Worklist.insert(AI);
3711 else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I))
3712 if (auto AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress()))
3713 DbgDeclares.insert(std::make_pair(AI, DDI));
3716 bool Changed = false;
3717 // A set of deleted alloca instruction pointers which should be removed from
3718 // the list of promotable allocas.
3719 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3722 while (!Worklist.empty()) {
3723 Changed |= runOnAlloca(*Worklist.pop_back_val());
3724 deleteDeadInstructions(DeletedAllocas);
3726 // Remove the deleted allocas from various lists so that we don't try to
3727 // continue processing them.
3728 if (!DeletedAllocas.empty()) {
3729 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
3730 Worklist.remove_if(IsInSet);
3731 PostPromotionWorklist.remove_if(IsInSet);
3732 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3733 PromotableAllocas.end(),
3735 PromotableAllocas.end());
3736 DeletedAllocas.clear();
3740 Changed |= promoteAllocas(F);
3742 Worklist = PostPromotionWorklist;
3743 PostPromotionWorklist.clear();
3744 } while (!Worklist.empty());
3749 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3750 AU.addRequired<AssumptionTracker>();
3751 if (RequiresDomTree)
3752 AU.addRequired<DominatorTreeWrapperPass>();
3753 AU.setPreservesCFG();